Use Skills

Helps choose relevant installed skills for complex requests and briefly shows the working set before answering.

Use this skill when the user invokes $use-skills, or when a request clearly benefits from combining more than one installed skill.

Purpose

Choose a small, relevant working set of skills and turn their guidance into one coherent result.

Good fits include requests that combine areas such as:

• planning and implementation
• review and testing
• writing and structure
• design and frontend work
• documentation and code changes

Skip this skill when one domain skill is clearly enough.

Selection Process

0. Resolve the skill mode using the hard gate below.
1. Review the visible skill list.
2. Compare each skill with the user request and expected output.
3. Pick the strongest matches as the working set.
4. Use supporting skills only when they materially improve the result.
5. Read only the parts of selected skills that help with the current task.
6. Produce one unified answer, plan, patch, or recommendation.

Keep the working set appropriate to the selected mode. Recommended is the usual choice when the user wants balanced quality, but ask for a mode when no reusable prior choice exists.

Match Labels

Use these labels while choosing:

• primary: directly shapes the result
• support: adds useful quality, structure, review, wording, or edge-case coverage
• skip: not useful for the current request

If no skill is a strong fit, leave this skill unused.

Modes

Ask the user to choose one mode before any workspace exploration, tool calls, file reads, skill selection, or main response when there is no reusable prior choice:

• All related: use every available skill that is meaningfully related to the request
• Recommended: use the best balanced working set for the request
• Restricted: use only the strongest matches, usually one to three skills

Ask with this compact terminal-friendly format, inside a fenced text code block:

1. All related - use every available skill that is meaningfully related.
   Using: use-skills, <all related skill candidates>
   For: broad coverage across <purposes>

2. Recommended - use the best balanced working set.
   Using: use-skills, <recommended skill candidates>
   For: strong output without unnecessary noise

3. Restricted - use only the strongest matches.
   Using: use-skills, <one to three strongest skill candidates>
   For: focused output with minimal skill involvement

Choose skill mode. Reply with 1, 2, or 3.

Mode Choice Hard Gate

If $use-skills is invoked and the prompt does not explicitly name All related, Recommended, or Restricted, the next assistant response must be only:

the fenced text mode menu above.

Do not inspect files, search the workspace, read skill files, or infer a final mode before asking.

The menu should still mention likely skills for each option. Use only the current prompt and visible/provided skill names/descriptions to draft those candidate lists. Do not use tools to discover more context before the user chooses.

Candidate sources allowed before mode choice:

• installed skill metadata already visible in the current session
• skill blocks pasted or provided by the user in the current conversation, such as <skill><name>enhance-prompt</name>...
• skills explicitly named by the user with $skill-name, when a visible/provided description exists in the session

Replace the placeholder skill candidates with actual visible/provided skill names whenever they are available. In the mode menu, use bare skill names such as use-skills, brainstorming, and writing-plans without a $ prefix. If the visible skill list is unavailable, say skills selected after mode choice instead of inventing names.

For All related, be aggressive. Include every candidate with a meaningful primary, support, adjacent-context, prompt-quality, wording, planning, or framing role. If a skill is commonly helpful for making the prompt, context, plan, or output better, include it in All related even when it is not the narrowest domain match. If the user explicitly names or provides a skill and it has any plausible support role, include it in All related even when it is too broad or weak for Recommended. Example: include enhance-prompt when the task involves improving prompts, examples, mode menus, docs, UI prompts, or prompt-facing wording.

When the user asks to improve a prompt, tighten a prompt, clarify prompt wording, make prompt/context better, or includes prompt-facing wording like Patch this bug report so it is clearer..., include enhance-prompt in All related if it is installed, visible, provided in the conversation, or explicitly mentioned. Do not exclude it from All related just because the current task is not a Stitch UI prompt; use it as support for prompt structure and clarity.

For Recommended, include brainstorming when the task needs strategy, context analysis, behavior changes, feature changes, prompt/context improvement, report framing, or deciding how to shape the work before execution. Recommended should usually include common support skills that materially improve output quality, not only the strictest domain skills.

The mode menu must be a standalone fenced text code block, not a Markdown bullet, not a paragraph, and not nested under another list item.

Spacing is mandatory:

• Insert one completely empty line after option 1's For: line before option 2.
• Insert one completely empty line after option 2's For: line before option 3.
• Insert one completely empty line after option 3's For: line before Choose skill mode. Reply with 1, 2, or 3.
• Do not add blank lines inside an option between the option title, Using:, and For:.

If 2. Recommended appears directly under option 1's For: line, or 3. Restricted appears directly under option 2's For: line, the menu is incorrectly formatted and must be rewritten with separator lines.

Do not use Markdown bold or all-caps labels in the mode menu because terminal transcripts may not render styling. Put Choose skill mode. Reply with 1, 2, or 3. after the three options, not before them. Do not add any explanatory sentence before or after the fenced block when asking for the mode.

Phrases like best, most relevant, strongest, helpful, or best combination do not count as explicit mode choices.

Do not choose silently unless the user already specified one of the three modes or a previous mode still applies. Reuse the previous mode only when the task and expected output are materially the same. Ask again when the task, expected output, or selected skill set changes.

Response Block

When this skill is used, start with a compact block:

• Mode: All related | Recommended | Restricted
• Using: use-skills, <selected skill>
• For: <short purpose>

List only skills that actually shape the answer. Keep selection details out of the response unless the user asks for them.

Conflict Handling

When selected skills point in different directions, prefer:

1. the user's current request
2. higher-priority session and project guidance
3. narrower domain skills over broader general skills
4. newer and more specific guidance over older or vague guidance

Quality Checks

• Keep the answer focused on the user's task.
• Avoid irrelevant skill advice.
• Avoid duplicate guidance.
• Keep the opening block shorter than the answer it introduces.
• Include a support skill in Using: only when it materially changes the result.
• Ask for mode before doing anything else when no reusable prior mode exists.
• Do not ask again when the prior mode still fits the task and expected output.
• Ask again when the task, expected output, or selected skill set has changed.

Example Triggers

• $use-skills
• Turn this feature idea into an implementation plan with testing notes.
• Rewrite this README so it is clearer and better structured.
• This spans planning, implementation, and review. Use the best combination of skills.
• Review this change and give the strongest findings first.
• Use restricted mode and only the strongest skills for this request.